Oil well yield prediction method based on multi-source information feature fusion
By using a multi-source information feature fusion method, combining displacement-load time series, dynamometer diagrams and production parameter data, and utilizing cross-attention mechanism and LSTM model, the problem of low oil well production prediction accuracy was solved, and more efficient prediction results were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEAST GASOLINEEUM UNIV
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-15
AI Technical Summary
Existing single prediction models for oil well production forecasting suffer from decreased accuracy and are prone to getting trapped in local optima as the application scope of oil fields increases and the amount of data grows. They also struggle to capture the deep-seated patterns of complex oil well production-related factors, resulting in low prediction accuracy.
A multi-source information feature fusion method is adopted. By collecting and preprocessing displacement-load time series data, dynamometer diagram data and production parameter data of oil wells, feature fusion and dimensionality reduction are performed using cross-attention mechanism and LSTM model to construct an integrated model for oil well production prediction.
It improves the accuracy of oil well production forecasting and the computational efficiency of the model, enabling it to more accurately capture the deep-seated patterns of complex oil well production-related factors and enhance oil well production efficiency.
Smart Images
Figure CN122046239A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oil and gas development technology, and more specifically to a method for predicting oil well production through multi-source information feature fusion. Background Technology
[0002] As a fundamental energy source for global economic and social development, oil is directly related to national economic stability and national defense security, and is a core support of the modern industrial system. Oil well production forecasting, as a core task of oilfield management, determines the efficient development and rational utilization of oil resources. Especially against the backdrop of continuously growing global oil demand, ensuring stable oil well production has become crucial for maintaining energy security.
[0003] Oil well production forecasting is a crucial aspect of oilfield development, used to reflect real-time reservoir dynamics. Accurately predicting production volume is a vital task for oilfield management to ensure stable well output. Early production forecasting relied on traditional empirical formulas, which were convenient and easy to implement, but had low accuracy. In recent years, artificial intelligence has been widely applied in oilfield development and production. Through training and analysis of large amounts of historical production sample data, it has achieved higher prediction accuracy.
[0004] However, with the increasing application scope of oilfields, the lengthening of prediction time spans, and the continuous increase in data volume, traditional single prediction models not only suffer from reduced accuracy but also easily get trapped in local optima, affecting prediction performance. Therefore, to address the production prediction needs of high water-cut, low-productivity oil wells, this paper proposes an intelligent prediction method that integrates multi-source information features and multi-model integration. This is an effective way to solve the problem of low accuracy in intelligent production prediction in current oil well development and a technical challenge that urgently needs to be overcome by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides an oil well production prediction method based on multi-source information feature fusion, which solves the problems existing in the background technology.
[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting oil well production through multi-source information feature fusion includes the following steps: S1. Collect displacement-load time-series data, dynamometer diagram data, and production parameter data during the oil well production process to form a multi-source information dataset; S2. Perform outlier handling, missing value imputation, and data standardization on the multi-source information dataset to obtain a high-quality preprocessed dataset. S3. The preprocessed high-quality dataset is divided into displacement-load features, dynamometer shape features and production parameter features according to feature type. For each type of feature, a matching dimensionality reduction method is used to remove redundancy and compress dimensions. S4. Using the displacement-load characteristics after dimensionality reduction as the main feature, the shape features of the dynamometer diagram and the production parameter features after dimensionality reduction are sequentially fused through the cross-attention mechanism to obtain the preliminary fused features; S5. Input the preliminary fused features into the fully connected layer for dimensionality compression, output the final fused features, and then input the final fused features into the LSTM model for training and validation to obtain the trained oil well production prediction ensemble model. S6. Using the trained integrated model for oil well production prediction, obtain the production prediction results for the oil well to be predicted.
[0007] Optionally, in S1, the displacement-load time sequence data is the displacement and load time sequence collected during the operation of the pumping unit, the indicator diagram graphic data is the closed curve image data drawn by the indicator instrument, and the production parameter data includes at least one of oil pressure, casing pressure, water cut, stroke, stroke frequency, pump depth, dynamic fluid level, submersion degree, current, torque, and power.
[0008] Optionally, the specific steps of S2 are as follows: S21. Adopt the 3Sigma criterion, based on the data mean. and standard deviation Identify outliers and remove those that meet the criteria. Data points; S22. For missing numerical data, the mean imputation method is used to fill in the missing data. S23. Using the Min-Max normalization method, all features are mapped to the [0,1] interval.
[0009] Optionally, the specific steps of S3 are as follows: S31. A 1D convolutional autoencoder with a multi-branch fusion embedded attention mechanism is used to reduce the dimensionality of displacement-load features to obtain 36-dimensional displacement-load dimensionality-reduced features. The 1D convolutional autoencoder includes 4 feature extraction modules, each layer has 2 convolutional kernel branches of different scales and 2 cross-layer fusion branches. All convolutional layers use the LeakyReLU activation function and remove pooling layers. S32. Binarize the indicator map, perform Canny edge detection and edge refinement, extract the 7-dimensional indicator map shape features by Hu invariant moments, and use supervised UMAP to obtain the 2-dimensional indicator map shape dimensionality reduction features. S33. Redundant feature pairs with an absolute value greater than 0.7 are removed by using the Pearson correlation coefficient, 5 redundant features are removed, and then the Boruta algorithm is used to select 7-dimensional production parameter dimensionality reduction features.
[0010] Optionally, in S31, the processing flow of the embedded attention mechanism is as follows: Global average pooling and global max pooling are performed on the displacement-load features after multi-branch fusion to obtain two C×1×1 feature vectors, which are then reduced to C8×1×1 via a 1×1 convolutional layer; the weights are pre-adjusted through two SE sub-modules, and the pre-adjusted channels are input into the 1×1 convolutional layer. The calculation formula is as follows:
[0011] In the formula: It is the sigmoid activation function. , These are weighted feature maps obtained from different branches. This is the output feature vector of global average pooling. This is the output feature vector of global max pooling.
[0012] Optionally, in S32, the correlation between the 2D dynamometer shape feature obtained by supervised UMAP dimensionality reduction and daily oil production is higher than that of the 7D dynamometer shape feature extracted by Hu invariant moments before dimensionality reduction.
[0013] Optionally, in S33, the Boruta algorithm generates 15 sets of random shadow features and compares them with the Z-score of the original features to select 6 important features with Z-scores higher than the baseline and 1 potential important feature. Among them, the important features include water content, pump depth, measured strokes, oil pressure, theoretical displacement, and measured stroke, and the potential important feature is the minimum load.
[0014] Optionally, the specific steps of S4 are as follows: S41. Using the 36-dimensional displacement-load dimension reduction feature as query Q and the 2-dimensional dynamometer diagram shape dimension reduction feature as key K and value V, a linear transformation matrix is applied... , , Mapped to the same space; according to the formula Attention weights are calculated and weighted summation is performed. Combined with residual connections and layer normalization, a 36-dimensional first fusion feature is obtained; among which, The dimension of key K; S42. Using the 36-dimensional first fusion feature as query Q and the 7-dimensional production parameter dimensionality reduction feature as key K and value V, repeat the above linear transformation, attention calculation and normalization steps to obtain the 36-dimensional preliminary fusion feature.
[0015] Optionally, in S5, the fully connected layer compresses the 36-dimensional initial fused features to 12 dimensions through linear transformation, and then uses the LeakyReLU activation function for nonlinear processing to output the final fused features.
[0016] As can be seen from the above technical solution, compared with the prior art, this invention discloses a multi-source information feature fusion method for oil well production prediction, which can solve the problem of low accuracy in oil well production prediction due to the influence of multiple factors. Traditional single prediction methods are prone to getting trapped in local optima, and due to the high complexity of oil well production-related factors, it is difficult to capture deep-seated patterns in the data. However, this invention can achieve more accurate production prediction through multi-source information fusion, providing stronger support and solutions for improving oil well production efficiency. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0018] Figure 1 Architecture diagram of the oil well production prediction method based on multi-source information feature fusion provided by the present invention; Figure 2 This invention provides a multi-branch feature fusion structure based on an attention mechanism. Figure 3 The overall framework diagram of the CAE model structure based on multi-branch fusion ESE provided for this invention; Figure 4 This is a diagram illustrating the indicator diagram processing procedure provided by the present invention. Figure 5 This is a comparison chart of the fitting degree of the original splicing features, dimensionality-reduced splicing features, and dimensionality-reduced fusion features provided by the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] This invention discloses a method for predicting oil well production by fusing multi-source information features, such as... Figure 1 As shown, it includes the following steps: S1. Multi-source oil well data acquisition: Collect displacement-load time series data, dynamometer diagram data and production parameter data during the oil well production process to form a multi-source information dataset; S2. Data preprocessing: Outlier handling, missing value imputation, and data standardization are performed on the multi-source information dataset to obtain a high-quality preprocessed dataset. S3. Feature Classification and Targeted Dimensionality Reduction: The preprocessed high-quality dataset is divided into displacement-load features, dynamometer diagram shape features, and production parameter features according to feature type. For each type of feature, a matching dimensionality reduction method is used to remove redundancy and compress dimensions. S4, such as Figure 2 Cross-attention mechanism feature fusion: taking the dimensionality-reduced displacement-load features as the main features, the dimensionality-reduced dynamometer shape features and production parameter features are fused sequentially through the cross-attention mechanism to obtain preliminary fused features; S5. Dimensionality reduction optimization of fully connected layer: The preliminary fused features are input into the fully connected layer for dimensionality compression, and the final fused features are output. The final fused features are then input into the LSTM model for training and validation to obtain the trained integrated model for oil well production prediction. S6. Oil well production prediction: Using the trained integrated model for oil well production prediction, obtain the production prediction results for the oil well to be predicted.
[0021] Furthermore, in S1 of this embodiment, the displacement-load time sequence data is the displacement and load time sequence collected during the operation of the pumping unit, the indicator diagram graphic data is the closed curve image data drawn by the indicator instrument, and the production parameter data includes at least one of oil pressure, casing pressure, water cut, stroke, stroke frequency, pump depth, dynamic fluid level, submersion degree, current, torque, and power.
[0022] Furthermore, in this embodiment, the specific steps of S2 are as follows: S21. Outlier Cleaning: Using the 3Sigma criterion, based on the data mean. and standard deviation Identify outliers and remove those that meet the criteria. The data points are cleaned to remove extreme outliers from the original data. S22. Missing Value Imputation: For numerical missing data, the mean imputation method is used, according to the formula... (in The characteristic mean, n This represents the number of non-missing values. Imputation is performed on non-missing values, and after imputation, the proportion of missing values for each feature is reduced to 0; S23. Data Standardization: Adopt the Min-Max standardization method, according to the formula... (in , (These are the maximum and minimum values of the features, respectively) Map all features to the interval [0,1] to eliminate dimensional differences.
[0023] Furthermore, in this embodiment, the specific steps of S3 are as follows: S31, Dimensionality reduction of displacement-load characteristics: such as Figure 3 A 1D convolutional autoencoder (CAE) with a multi-branch fusion embedded attention mechanism (ESE) is used to reduce the dimensionality of displacement-load features, resulting in 36-dimensional displacement-load dimensionality-reduced features. The 1D convolutional autoencoder includes four feature extraction modules, each with two convolutional kernel branches of different scales (13×1 and 17×1) and two cross-layer fusion branches. All convolutional layers use the LeakyReLU activation function and pooling layers are removed. S32. Dimensionality reduction of indicator diagram shape features: such as... Figure 4 The indicator map is binarized (blue pixel determination threshold: B>0.65(R+G), where R, G, and B are the pixel values of the red, green, and blue channels, respectively), Canny edge detection (Gaussian filter standard deviation) and edge thinning are performed. The 7-dimensional indicator map shape features are extracted by Hu invariant moments, and the 2-dimensional indicator map shape dimensionality reduction features are obtained by using supervised UMAP (with daily oil production as the label). The formula for calculating Gaussian filtering is as follows:
[0024] S33. Dimensionality reduction of production parameter features: First, redundant feature pairs with an absolute value greater than 0.7 (such as dynamic liquid level-submersion degree (r=0.969) and downflow current-upflow current (r=0.738)) are removed by using the Pearson correlation coefficient. Five redundant features are removed. Then, the Boruta algorithm is used to select seven dimensions of production parameter dimensionality reduction features to provide an optimized feature dataset for subsequent prediction models.
[0025] Furthermore, in S31, the processing flow of the embedded attention mechanism (ESE) is as follows: Global average pooling (GAP) and global max pooling (GMP) are performed on the displacement-load features (dimension C×H×1) after multi-branch fusion, respectively, to obtain two C×1×1 feature vectors. These vectors are then reduced to C8×1×1 via a 1×1 convolutional layer (scaling factor R=8). The weights are pre-adjusted using two SE sub-modules (scaling factor = 2), and the pre-adjusted channels are input into a 1×1 convolutional layer to further optimize channel weight allocation and achieve feature enhancement. The calculation formula is:
[0026] In the formula: It is the sigmoid activation function. , These are weighted feature maps obtained from different branches. This is the output feature vector of global average pooling. This is the output feature vector of global max pooling.
[0027] Furthermore, in S32, the correlation between the 2D dynamometer shape feature obtained by supervised UMAP dimensionality reduction and daily oil production is higher than that of the 7D dynamometer shape feature extracted by Hu invariant moments before dimensionality reduction.
[0028] Seven invariant moments (f1-f7) are typically extracted from Hu invariant moments. f1 is scale-related, while f2-f7 are more complex features. The Pearson correlation coefficients of f2, f3, and f4 with daily oil production are 0.3124 (weak positive correlation), 0.3442 (weak positive correlation), and 0.2407 (weaker positive correlation), respectively, with p < 0.01. The Pearson correlation coefficients of the dimensionality reduction features (PF1, PF2) of the 2D dynamometer map obtained by supervised UMAP with daily oil production are -0.804 and 0.339, respectively (p < 0.01).
[0029] The formula for the Pearson correlation coefficient is as follows:
[0030] In the formula: x and y These are the sample values of the two variables, and These are their respective averages.
[0031] Furthermore, in S33, the Boruta algorithm generates 15 sets of random shadow features and compares them with the Z-score of the original features. It then selects 6 important features and 1 potentially important feature whose Z-score is higher than the baseline (ShadowMax). After selection, the average correlation coefficient between the features and the output increases to 0.32. Among them, the important features include water content, pump depth, measured strokes, oil pressure, theoretical displacement, and measured stroke, while the potentially important feature is the minimum load.
[0032] Furthermore, in this embodiment, the specific steps of S4 are as follows: S41. First Fusion: Using the 36-dimensional displacement-load dimensionality reduction feature as query Q and the 2-dimensional dynamometer diagram shape dimensionality reduction feature as key K and value V, a linear transformation matrix is applied... , , Mapped to the same space; according to the formula Attention weights are calculated and weighted summation is performed. Combined with residual connections and layer normalization, a 36-dimensional first fusion feature is obtained; among which, Let K be the dimension of the key.
[0033] Specifically, the displacement-load features (36-dimensional) and dynamometer diagram shape features (2-dimensional) are first fused. Since the displacement-load features are the primary features, they act as the query (Q), actively focusing on the key information contained within the dynamometer diagram shape features. The shape features, in turn, serve as the key (K) and value (V), providing reference information for feature enhancement. The dynamometer diagram shape features are dimensionality-reduced using the Supervised UMAP method, with well production as the dimensionality reduction label, ensuring that the extracted low-dimensional features are highly correlated with production. Through a cross-attention mechanism, the displacement-load features can adaptively fuse local detail information from the dynamometer diagram shape features, effectively supplementing the expressive power of morphological information in the load features and enhancing the displacement-load features' ability to perceive the shape of the dynamometer diagram. Ultimately, the dimensionality of the fused displacement-load and dynamometer diagram shape features remains 36-dimensional.
[0034] S42. Second fusion: Using the 36-dimensional first fusion feature as query Q and the 7-dimensional production parameter dimensionality reduction feature as key K and value V, repeat the above linear transformation, attention calculation and normalization steps to obtain the 36-dimensional preliminary fusion feature.
[0035] Furthermore, in S5 of this embodiment, the fully connected layer compresses the 36-dimensional preliminary fused features to 12 dimensions through linear transformation, and then uses the LeakyReLU activation function for nonlinear processing to output the final fused features, avoiding the problem of dead neurons. This step can improve the training efficiency and prediction performance of the model while reducing computational overhead.
[0036] Specifically, the fully connected layer maps the input features to the output feature space through a weighted and biased linear transformation, as shown in the following formula:
[0037] In the formula: x Given the input vector, W This is the weight matrix. b For bias terms, y This is the vector after dimensionality reduction.
[0038] Use the LeakyReLU activation function to apply the linearly transformed features y A nonlinear transformation is performed to enable the fully connected layer to extract more complex features. The calculation formula is as follows:
[0039] In the formula: a It should be a very small positive number to avoid the problem of dead neurons; ZThese are the final dimensionality-reduced features. The dimensionality of the fused multi-source features after dimensionality reduction through a fully connected layer is 12 dimensions.
[0040] The model is trained according to steps S1, S2, S3, S4, and S5. Figure 5 The comparison results show that feature reduction and fusion significantly optimize the data structure, remove redundant information, and retain key information highly relevant to output. Feature fusion not only enhances the representation of useful information but also improves the computational efficiency of the model, making it more widely applicable and efficient in practical applications. With the addition of dimensionality reduction and feature fusion, the predicted values gradually move closer to the fitted red line. y=x Closer proximity and denser distribution indicate improved prediction accuracy. In conclusion, multi-source feature fusion provides an effective solution for improving prediction accuracy and computational efficiency, verifying that a reasonable feature processing strategy plays a crucial role in optimizing model performance during large-scale data processing.
[0041] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting oil well production through multi-source information feature fusion, characterized in that, Includes the following steps: S1. Collect displacement-load time-series data, dynamometer diagram data, and production parameter data during the oil well production process to form a multi-source information dataset; S2. Perform outlier handling, missing value imputation, and data standardization on the multi-source information dataset to obtain a high-quality preprocessed dataset. S3. The preprocessed high-quality dataset is divided into displacement-load features, dynamometer shape features and production parameter features according to feature type. For each type of feature, a matching dimensionality reduction method is used to remove redundancy and compress dimensions. S4. Using the displacement-load characteristics after dimensionality reduction as the main feature, the shape features of the dynamometer diagram and the production parameter features after dimensionality reduction are sequentially fused through the cross-attention mechanism to obtain the preliminary fused features; S5. Input the preliminary fused features into the fully connected layer for dimensionality compression, output the final fused features, and then input the final fused features into the LSTM model for training and validation to obtain the trained oil well production prediction ensemble model. S6. Using the trained integrated model for oil well production prediction, obtain the production prediction results for the oil well to be predicted.
2. The oil well production prediction method based on multi-source information feature fusion according to claim 1, characterized in that, In S1, the displacement-load time sequence data is the displacement and load time sequence collected during the operation of the pumping unit, the indicator diagram data is the closed curve image data drawn by the indicator instrument, and the production parameter data includes at least one of the following: oil pressure, casing pressure, water cut, stroke, stroke frequency, pump depth, dynamic liquid level, submersion degree, current, torque, and power.
3. The oil well production prediction method based on multi-source information feature fusion according to claim 1, characterized in that, The specific steps of S2 are as follows: S21. Adopt the 3Sigma criterion, based on the data mean. and standard deviation Identify outliers and remove those that meet the criteria. Data points; S22. For missing numerical data, the mean imputation method is used to fill in the missing data. S23. Using the Min-Max normalization method, all features are mapped to the [0,1] interval.
4. The oil well production prediction method based on multi-source information feature fusion according to claim 1, characterized in that, The specific steps for S3 are as follows: S31. A 1D convolutional autoencoder with a multi-branch fusion embedded attention mechanism is used to reduce the dimensionality of displacement-load features to obtain 36-dimensional displacement-load dimensionality-reduced features. The 1D convolutional autoencoder includes 4 feature extraction modules, each layer has 2 convolutional kernel branches of different scales and 2 cross-layer fusion branches. All convolutional layers use the LeakyReLU activation function and remove pooling layers. S32. Binarize the indicator map, perform Canny edge detection and edge refinement, extract the 7-dimensional indicator map shape features by Hu invariant moments, and use supervised UMAP to obtain the 2-dimensional indicator map shape dimensionality reduction features. S33. Redundant feature pairs with an absolute value greater than 0.7 are removed by using the Pearson correlation coefficient, 5 redundant features are removed, and then the Boruta algorithm is used to select 7-dimensional production parameter dimensionality reduction features.
5. The oil well production prediction method based on multi-source information feature fusion according to claim 4, characterized in that, In S31, the embedded attention mechanism's processing flow is as follows: Global average pooling and global max pooling are performed on the displacement-load features after multi-branch fusion to obtain two C×1×1 feature vectors, which are then reduced to C8×1×1 via a 1×1 convolutional layer. The weights are pre-adjusted through two SE sub-modules, and the pre-adjusted channels are input into the 1×1 convolutional layer. The calculation formula is as follows: In the formula: It is the sigmoid activation function. , These are weighted feature maps obtained from different branches. This is the output feature vector of global average pooling. This is the output feature vector of global max pooling.
6. The oil well production prediction method based on multi-source information feature fusion according to claim 4, characterized in that, In S32, the correlation between the 2D dynamometer shape feature obtained by supervised UMAP dimensionality reduction and daily oil production is higher than that of the 7D dynamometer shape feature extracted by Hu invariant moments before dimensionality reduction.
7. The oil well production prediction method based on multi-source information feature fusion according to claim 4, characterized in that, In S33, the Boruta algorithm generates 15 sets of random shadow features and compares them with the Z-score of the original features to select 6 important features with Z-scores higher than the baseline and 1 potential important feature. Among them, the important features include water content, pump depth, measured strokes, oil pressure, theoretical displacement, and measured stroke, and the potential important feature is the minimum load.
8. The oil well production prediction method based on multi-source information feature fusion according to claim 4, characterized in that, The specific steps of S4 are as follows: S41. Using the 36-dimensional displacement-load dimension reduction feature as query Q and the 2-dimensional dynamometer diagram shape dimension reduction feature as key K and value V, a linear transformation matrix is applied... , , Mapped to the same space; according to the formula Attention weights are calculated and weighted summation is performed. Combined with residual connections and layer normalization, a 36-dimensional first fusion feature is obtained; among which, The dimension of key K; S42. Using the 36-dimensional first fusion feature as query Q and the 7-dimensional production parameter dimensionality reduction feature as key K and value V, repeat the above linear transformation, attention calculation and normalization steps to obtain the 36-dimensional preliminary fusion feature.
9. The oil well production prediction method based on multi-source information feature fusion according to claim 8, characterized in that, In S5, the fully connected layer compresses the 36-dimensional initial fused features to 12 dimensions through linear transformation, and then uses the LeakyReLU activation function for nonlinear processing to output the final fused features.